分类： 物理学 >> 核物理学 提交时间： 2025-04-01
摘要： Determining the release source position and quantity is crucial for evaluating the consequences of atmospheric radionuclide release events, with the Bayesian method serving as the primary tool for source inversion. Reducing the impact of input data errors on inversion uncertainty and improving computational efficiency are key to developing robust and efficient inversion algorithms. To address these challenges, a spatiotemporal trajectory prior (STP) distribution was developed to effectively mitigate the influence of measurement and simulation errors on inversion results without increasing computational costs, thereby enhancing the robustness and accuracy of the inversion process. Additionally, a joint adaptive Markov Chain Monte Carlo (MCMC) sampling method was introduced, integrating the traditional parallel tempering (PT) algorithm with a novel joint adaptive transition proposal (JATP) algorithm to accelerate inversion calculations. The proposed methods were optimized and validated using data from the first release of the European Tracer Experiment (ETEX-I). After determining the hyperparameters, the JATP algorithm consistently maintained the sampling process near the theoretically optimal acceptance rate of 0.234. The PT algorithm, utilizing an optimized temperature schedule, achieved a 3.3-fold improvement in sampling efficiency compared to single-chain sampling. Under bootstrap statistical comparison, the method reduced the relative error of position, relative error of release quantity, and combined relative error by 25.9%, 27.7%, and 27.8%, compared to the traditional uniform prior method, respectively. And the deviation of the estimated and true source position is within 0.25˚. The results demonstrate the accuracy and effectiveness of our method.
分类： 物理学 >> 核物理学 提交时间： 2025-02-08
Yan, Dr. Shu-chang
QIU, Dr. RUI 邱睿
Hu, Dr. Ankang
Qu, Dr. Shuiyin
Zhou, Dr. Yang
Hu, Dr. Zi-yi
Zhou, Dr. Yanhan
Wu, Dr. Zhen
LI, Prof. JUNLI
摘要： Mesh-type phantoms represent the latest generation of human computational phantoms, offering high resolution and adjustability advantages for individualized radiation dosimetry. Current dosimetry computation methods, which require conversion to tetrahedral mesh models for efficient Monte Carlo simulations, still do not meet the requirements for real-time dose calculations. Advancements in heterogeneous computing now allow for significant acceleration in mesh-type phantom calculations by utilizing both high-performance hardware and efficient algorithms. This study aims to develop a GPU-accelerated Monte Carlo simulation method that directly utilizes mesh-type phantoms to further enhance the speed of human dose calculations without the need for tetrahedralization. For the boundary representation polygonal models, this study redesigned and implemented the entire procedural flow of the GPU-accelerated Monte Carlo program, developing particle transport methods within the mesh-type model. All triangular facets of the mesh-type model were constructed into a tree-like acceleration structure and the traversal access pattern was optimized. Moreover, this study adopted an event-based transport method, transporting particles step-by-step by particle type, and a bias-based variance reduction technique employing geometric weights was integrated. For typical external irradiation scenarios, dose calculations between Geant4 and our GPU-based program were compared to assess computational accuracy and efficiency. Compared to the benchmark simulations conducted on a single-thread CPU via Geant4, the organ dose discrepancies calculated by the GPU-accelerated program generally remained within a 5% margin, while computational times were reduced by factors ranging from 500 to 50000. To our knowledge, this study is the first to utilize a mesh-type model for GPU-accelerated dose calculation without tetrahedralization. The simulation time has been dramatically reduced from hours to just mere seconds, offering a rapid and precise Monte Carlo method for mesh-type computational phantoms. This development supports real-time dose calculation studies using dynamic mesh-type models, providing a robust Monte Carlo simulation tool.
分类： 物理学 >> 核物理学 提交时间： 2025-01-02
摘要： Determining the release source position and quantity is crucial for evaluating the consequences of atmospheric radionuclide release events, with the Bayesian method serving as the primary tool for source inversion. Reducing the impact of input data errors on inversion uncertainty and improving computational efficiency are key to developing robust and efficient inversion algorithms. To address these challenges, we developed a spatiotemporal trajectory prior (STP) distribution that effectively mitigates the influence of measurement and simulation errors on inversion results without increasing computational costs, thereby enhancing the robustness and accuracy of the inversion process. Additionally, we introduced a joint adaptive Markov Chain Monte Carlo (MCMC) sampling method that integrates the traditional parallel tempering (PT) algorithm with a novel joint adaptive transition proposal (JATP) algorithm to accelerate inversion calculations. The proposed methods were optimized and validated using data from the first release of the European Tracer Experiment (ETEX-I). After determining the hyperparameters, the JATP algorithm consistently maintained the sampling process near the theoretically optimal acceptance rate of 0.234. The PT algorithm, utilizing an optimized temperature schedule, achieved a 2.89-fold improvement in sampling efficiency compared to single-chain sampling. Under bootstrap statistical comparison, the method reduced the relative error of position, relative error of release quantity, and total relative error by 25.9%, 27.7%, and 27.8%, compared to the traditional uniform prior method, respectively. And the deviation of the estimated and true source position is within 0.25˚. The results demonstrate the accuracy and effectiveness of our method.
分类： 物理学 >> 核物理学 提交时间： 2024-12-16
Yan, Dr. Shu-chang
QIU, Dr. RUI 邱睿
Hu, Dr. Ankang
Qu, Dr. Shuiyin
Zhou, Dr. Yang
Hu, Dr. Zi-yi
Zhou, Dr. Yanhan
Wu, Dr. Zhen
LI, Prof. JUNLI
摘要： Mesh-type phantoms represent the latest generation of human computational phantoms, offering high resolution and adjustability advantages for individualized radiation dosimetry. Current dosimetry computation methods, which require conversion to tetrahedral mesh models for efficient Monte Carlo simulations, still do not meet the requirements for real-time dose calculations. Advancements in heterogeneous computing now allow for significant acceleration in mesh-type phantom calculations by utilizing both high-performance hardware and efficient algorithms. This study aims to develop a GPU-accelerated Monte Carlo simulation method that directly utilizes mesh-type phantoms to further enhance the speed of human dose calculations without the need for tetrahedralization. For the boundary representation polygonal models, this study redesigned and implemented the entire procedural flow of the GPU-accelerated Monte Carlo program, developing particle transport methods within the mesh-type model. All triangular facets of the mesh-type model were constructed into a tree-like acceleration structure and the traversal access pattern was optimized. Moreover, this study adopted an event-based transport method, transporting particles step-by-step by particle type, and a bias-based variance reduction technique employing geometric weights was integrated. For typical external irradiation scenarios, dose calculations between Geant4 and our GPU-based program were compared to assess computational accuracy and efficiency. Compared to the benchmark simulations conducted on a single-thread CPU via Geant4, the organ dose discrepancies calculated by the GPU-accelerated program generally remained within a 5% margin, while computational times were reduced by factors ranging from 500 to 50000. To our knowledge, this study is the first to utilize a mesh-type model for GPU-accelerated dose calculation without tetrahedralization. The simulation time has been dramatically reduced from hours to just mere seconds, offering a rapid and precise Monte Carlo method for mesh-type computational phantoms. This development supports real-time dose calculation studies using dynamic mesh-type models, providing a robust Monte Carlo simulation tool.
分类： 物理学 >> 核物理学 提交时间： 2024-12-16
WU, Dr. Jiajun
Zhang, Dr. Hui
HU, Dr. Ankang
PU, Mr. Yanheng
Wu, Dr. Zhen
QIU, Dr. RUI 邱睿
LI, Prof. JUNLI
摘要： This paper introduces a novel algorithm aimed at enhancing the computational performance of Monte Carlo codes: A Dynamic Adaptive Load Balancing Algorithm for Monte Carlo Codes (DALB-MC algorithm). With the widespread use of multicore processors, some Monte Carlo codes have been adapted to leverage both local thread parallelism and global parallelism through message passing between nodes, enabling concurrent execution of independent tasks and efficient merging of results. However, as shielding engineering problems increase in complexity and computational environments become more heterogeneous, traditional parallel computing algorithms face limitations in handling Monte Carlo simulations, particularly in ensuring adapting to diverse hardware resources. The proposed DALB-MC algorithm addresses these challenges by dynamically partitioning tasks into multiple sub-tasks and allocating them in real-time based on the computational capacity of each node, thereby optimizing resource utilization and reducing overall simulation time. In this study, we implement DALB-MC algorithm on MCShield and demonstrate the correctness and effectiveness of the algorithm in improving the performance of the Monte Carlo code through self-defined arithmetic examples. The experimental results show that compared with original algorithm, DALB-MC algorithm achieves about 20% computation time acceleration in massively parallel computation and significantly improves the overall efficiency of the program. In addition, a Linux supercomputing platform is used to test a complex benchmark case to further verify the optimization effect of DALB-MC algorithm under the computational scale of real engineering applications.
点击量
待评议
点击量
下载量
评论
